Hebbian Deep Learning Without Feedback

Authors: Adrien Journé, Hector Garcia Rodriguez, Qinghai Guo, Timoleon Moraitis

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental accuracies on MNIST, CIFAR-10, STL-10, and Image Net, respectively reach 99.4%, 80.3%, 76.2%, and 27.3%. In conclusion, Soft Hebb shows with a radically different approach from BP that Deep Learning over few layers may be plausible in the brain and increases the accuracy of bio-plausible machine learning. Code is available at https://github.com/Neuromorphic Computing/Soft Hebb.
Researcher Affiliation Industry Adrien Journ e1, Hector Garcia Rodriguez1, Qinghai Guo2, Timoleon Moraitis1* {adrien.journe, hector.garcia.rodriguez, guoqinghai, timoleon.moraitis}@huawei.com 1Huawei Zurich Research Center, Switzerland 2Huawei ACS Lab, Shenzhen, China
Pseudocode No The paper describes the Soft Hebb algorithm and its plasticity rule using equations, but it does not include pseudocode or a clearly labeled algorithm block.
Open Source Code Yes Code is available at https://github.com/Neuromorphic Computing/Soft Hebb.
Open Datasets Yes accuracies on MNIST, CIFAR-10, STL-10, and Image Net, respectively reach 99.4%, 80.3%, 76.2%, and 27.3%.
Dataset Splits Yes All grid searches were performed on three different random seeds, varying the batch sampling and the validation set (20% of the training dataset).
Hardware Specification Yes We used an NVIDIA Tesla V100 32GB GPU.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies used in the experiments. It only mentions "PyTorch" in the context of another paper's code (Appendix A.7).
Experiment Setup Yes The linear classifier on top uses a mini-batch of 64 and trains on 50 epochs for MNIST and CIFAR-10, 100 epochs for STL-10, and 200 epochs for Image Net. For all datasets, the learning-rate has an initial value of 0.001 and is halved repeatedly at [20%, 35%, 50%, 60%, 70%, 80%, 90%] of the total number of epochs. Data augmentation (random cropping and flipping) was applied for STL-10 and Image Net.